Noise-Adaptive Regularization for Robust Multi-Label Remote Sensing Image Classification

TL;DR

Robust multi-label remote sensing image classification is challenged by additive and subtractive label noise arising from crowdsourced and thematic annotations. The paper introduces Noise-Adaptive Regularization (NAR), a semi-supervised framework that differentiates noise types using a three-state, confidence-based label handling, and couples it with a multi-label extension of Early-learning Regularization (ELR). Empirical results across three RS datasets and three noise scenarios show that NAR achieves state-of-the-art robustness, with the largest gains under subtractive and mixed noise, driven by adaptive supervision suppression and selective correction. This approach offers a practical, noise-aware strategy for improving reliability in RS MLC and suggests potential gains from integrating with self-supervised signals in future work.

Abstract

The development of reliable methods for multi-label classification (MLC) has become a prominent research direction in remote sensing (RS). As the scale of RS data continues to expand, annotation procedures increasingly rely on thematic products or crowdsourced procedures to reduce the cost of manual annotation. While cost-effective, these strategies often introduce multi-label noise in the form of partially incorrect annotations. In MLC, label noise arises as additive noise, subtractive noise, or a combination of both in the form of mixed noise. Previous work has largely overlooked this distinction and commonly treats noisy annotations as supervised signals, lacking mechanisms that explicitly adapt learning behavior to different noise types. To address this limitation, we propose NAR, a noise-adaptive regularization method that explicitly distinguishes between additive and subtractive noise within a semi-supervised learning framework. NAR employs a confidence-based label handling mechanism that dynamically retains label entries with high confidence, temporarily deactivates entries with moderate confidence, and corrects low confidence entries via flipping. This selective attenuation of supervision is integrated with early-learning regularization (ELR) to stabilize training and mitigate overfitting to corrupted labels. Experiments across additive, subtractive, and mixed noise scenarios demonstrate that NAR consistently improves robustness compared with existing methods. Performance improvements are most pronounced under subtractive and mixed noise, indicating that adaptive suppression and selective correction of noisy supervision provide an effective strategy for noise robust learning in RS MLC.

Figures (5)

• Figure 1: Schematic illustration of the confidence-based label handling mechanism. Each label entry (i.e., 0 or 1) is handled via one of three states: (1) label entry retaining of likely clean entries, when $y_{i,c}$ is close to $p_{i,c}$, (2) label entry deactivation via setting weight $w_{i,c}$ to 0, when $p_{i,c}$ is neither close to 0 or 1, (3) label entry flipping when $y_{i,c}$ is close to $1 - p_{i,c}$. Resulting corrected label entries $\tilde{y}_{i,c} = 0$ if colored in blue, $\tilde{y}_{i,c} = 1$ if colored in orange, and deactivated when colored in beige.
• Figure 2: Example images of the datasets: \ref{['subfigure:example-ucmerced']} UCMerced, \ref{['subfigure:example-deepglobe']} DeepGlobe-ML, and \ref{['subfigure:example-aidml']} AID-ML.
• Figure 3: Example of inducing $20$% mixed noise (e.g., $20$% additive noise and 20% subtractive noise). $Y$ is the clean label matrix each row representing multi-label $y_i$, each column representing a class $c$. $\tilde{Y}$ represents the noisified label matrix. Additive noise is depicted in blue changing an entry $c$ in the label $y_i$ from $0$ to $1$, while subtractive noise is depicted in red changing an entry $c$ in the label $y_i$ from $1$ to $0$. Figure is reproduced from burgert_effects_2022.
• Figure 4: Oracle-based evaluation of label handling strategies under 40% label noise for UCMerced. Each cell shows the mAP macro for combinations of label handling applied to oracle-identified noisy label entries and an equal number of top $k$ most uncertain clean label entries per class. \ref{['subfigure:oracle_subn']} Subtractive noise. \ref{['subfigure:oracle_addn']} Additive noise.
• Figure 5: Parameter Sensitivity analysis for UCMerced. \ref{['subfigure:ablation_subn']} Subtractive noise from 10% to 60%. \ref{['subfigure:ablation_addn']} Additive noise from 10% to 60%. \ref{['subfigure:ablation_mixn']} Mixed noise for 40%.